Transformer Vibration Signal Analysis Based On Deep Learning

The power transformer is the pivotal equipment for electrical energy conversion in the power grid.For industrial manufacturing,residents’ life and social order,which is of great significance to ensure the safety and stable operation of my country’s power system.Therefore,it is necessary to monitor the transformer status in real-time and promptly detect latent faults in practical engineering applications.At present,the vibration analysis method has been widely studied due to its advantages of live online monitoring,direct response to the mechanical state,and no electrical connection.This thesis conducts research on analysis and identification method of transformer vibration signal based on deep learning.Firstly,the construction and transmission are studied.And the transformer vibration data for this thesis are presented.Furthermore,the frequency spectrum distribution of the vibration data at different measuring points of the transformer is analyzed.After that,through the simulation experiment,the ability to express the time domain and frequency domain features of the image after continuous wavelet transform is analyzed.Then,aiming at the problem of insufficient feature extraction ability of the deep neural network for the original vibration signal,a method of transformer vibration signal identification based on data segmentation is proposed.The method calculates the sample length and sliding step size based on the vibration cycle and sampling frequency.At the same time,to fully preserve the time correlation of time series vibration signals,an improved one-dimensional convolutional neural network model is established.Small convolution kernel,global average pooling,batch normalization layer,and dropout are introduced into the model to improve the fitting and generalization capability.The experimental results show that the proposed method can optimize the time domain segmentation of transformer vibration signals and significantly improve the feature extraction capability of the convolutional neural network for time-domain vibration data.Finally,aiming at the difficulty of feature extraction due to the influence of various factors and the high complexity of transformer vibration signals,which is difficult to be directly used for fault diagnosis,a vibration signal classification model based on a continuous wavelet time-frequency map was proposed.Each branch independently extracts different attribute features.In the CWUR dataset,the average fault identification accuracy of the model is over 99%,and it can diagnose unknown faults.In the transformer vibration dataset,the analysis results show that the average identification accuracy of the model in this thesis is over98%,and identified potential types.It provides a reference for transformer vibration fault detection under the condition of insufficient prior knowledge.